Support int32_t indices/offsets for caching handling logics (pytorch#811

)

Summary:
Pull Request resolved: pytorch#811

In training, we assume the indices / offsets are int64_t for embedding (TBE), but in inference, we assume the indices / offsets are int32_t.

This Diff enables both int32_t and int64_t supports for the caching logics so that we can reuse the same functions for both training and inference, while reducing the extra overhead to convert the indices/offsets from int to long or vice versa.

Differential Revision: D33045589

fbshipit-source-id: 42ebcd899bb5dc6735eaf67cad48ac3b168d60ca

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops.py

@@ -1824,7 +1824,6 @@ def prefetch(self, indices: Tensor, offsets: Tensor) -> None:
         if not self.lxu_cache_weights.numel():
             return
 
-        (indices, offsets) = indices.long(), offsets.long()
         linear_cache_indices = torch.ops.fb.linearize_cache_indices(
             self.cache_hash_size_cumsum,
             indices,